Re: POC, WIP: OR-clause support for indexes

On Fri, Oct 4, 2024 at 9:20 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> On Fri, Oct 4, 2024 at 2:00 PM Alexander Korotkov <aekorotkov(at)gmail(dot)com> wrote:
> > Yes, transformAExprIn() does the work to coerce all the expressions in
> > the right part to the same type. Similar logic could be implemented
> > in match_orclause_to_indexcol(). What worries me is whether it's
> > quite late stage for this kind of work. transformAExprIn() works
> > during parse stage, when we need to to resolve types, operators etc.
> > And we do that once.
>
> I agree that it would be a bit awkward. Especially having spent so
> much time talking about doing this later on, not during parsing. That
> doesn't mean that it's necessarily the wrong thing to do, though.
>
> > If we replicate the same logic to
> > match_orclause_to_indexcol(), then we may end up with index scan using
> > one operator and sequential scan using another operator.
>
> But that's already true today. For example, these two queries use
> different operators at runtime, assuming both use a B-Tree index scan:
>
> select * from tenk1 where four = any('{0,1}'::int[]) and four =
> any('{1,2}'::bigint[]);
>
> select * from tenk1 where four = any('{1,2}'::bigint[]) and four =
> any('{0,1}'::int[]); -- flip the order of the arrays, change nothing
> else
>
> This isn't apparent from what EXPLAIN ANALYZE output shows, but the
> fact is that only one operator (and one array) will be used at
> runtime, after nbtree preprocessing completes. I'm not entirely sure
> how this kind of difference might affect a sequential scan. I imagine
> that it can use either or both operators unpredictably.

Yes, but those operators are in the B-tree operator family. That
implies a lot about semantics of those operators making B-tree
legitimate to do such transformations. But it's different story when
you apply it to arbitrary operator and arbitrary implicit cast. I can
imagine implicit casts which could throw errors or loose precision.
It's OK to apply them as soon as user made them implicit. But
applying them in different ways for different optimizer decisions
looks risky.

------
Regards,
Alexander Korotkov
Supabase